A 2025 study on GPT-3 (a now outdated model, replaced by GPT-5) states that it "[consumes] a 500ml bottle of
water for roughly 10-50 medium-length responses, depending on when and where it is deployed."1
Remembering that this was a significantly less powerful model than the most recent, the chances of that number
increasing is likely; even then, the usage of water was concerning.
Water is measured through two terms: withdrawal and consumption. Withdrawal refers to how much total water is
taken from a source, and consumption refers to how much of that total was used throughout the process. Some is
used in cooling, most is used in generating the energy necessary for the servers2 - though it is
still difficult to track as many companies do not address nor publish their resource use.
1. (Li et al., 2025)
2. (Ren, 2023)
What are some positives?
AI could prove to be useful for quite a bit! Though not directly tied to the usage of water, management of water seems to be quite an effective use of AI. VODA.ai is one that sticks out - it's already been used in Tucson, Arizona since 2020.3 Its use is to predict when pipes may break, create a quarterly score on risk exposure amounts, evaluate data, and more.4 Along with VODA as an example, the Environmental Finance Center Network (EFCN) describes six major points - though sixth was repetitive, so one chose to remove it - that AI could help with in terms of water; a keyword, however, is can. It has the potential to help, and we have seen some already with VOMA, but most is still potential. The following points are all from the EFCN5:
1. Water quality management - in placing sensors in the water, AI can be used as a way to continuously measure
water quality. It's efficient as it takes less time than humans, and responses to changes in the water can be
caught much faster than relying on humans.
2. Leak detecting and prevention - similar to the one above, having a constant monitor for pipes that have/could
break is efficient. It can also sift through more data in a much quicker time than a human approach, so
potential leaks can be prevented as well.
3. Maintenance - specifically for infrastructure, AI can use data to change a lot of elements like pressure and
velocity in order to best use energy. The EFCN claims as well that some advanced systems may even have the
potential to prevent "detrimental sewage overflows during severe weather" which one could imagine to be useful
in places particularly susceptible to large storms.
4. Flood prediction - while focusing more on weather and less physical water management, predictive AI could be
useful for analyzing weather, river levels, and flood patterns to sound an early alarm bell on flooding.
5. Conservation - optimizing the amount of water used for tasks like irrigation by analyzing weather, soil, and
crop data will allow for less overall water usage. However, this is already something that we have, so perhaps
AI would simply be a step up in efficiency and reliability?
3. (Bonney, 2023)
4. (VODA.ai, n.d.)
5. (Bonney, 2023)
What are some negatives?
While AI shows promise to manage water, the amount of water it uses will become a danger far before that promise is viable for the majority, especially given how global warming is accelerating at a fast pace. Environmental impact of AI? More water is burned off, releasing more heat and pollution, causing more of the polar ice to melt - while it may give us more freshwater to use in these systems, those systems will not be particularly useful if they are underwater. As AI becomes more and more popular and demand for it grows, servers outfitted with multiple GPUs (Graphic processing units) will only rise; it just so happens that these servers, especially in large clusters, tend to be the most "hungry" for water and power, leading them to pollute the most as well. From the Organisation for Economic Co-operation and Development (OECD), specifically under their .ai label, describes more in detail the two places where LLMs use the most water6:
For scope-1 water consumption, the focus is onsite. This includes the water used in cooling; with a many servers
being incompatible with saltwater (for what one can assume to be either due to worry about interference with
electricity, salt buildup within the system, or a variety of other reasons) most systems use a significant
amount of freshwater, most of which is taken from the surrounding environment - be it groundwater or from a
water supply. While in theory, much of this could be done with a closed-system loop where no water escapes, that
is not something viable either - as the water needs to be cooled back down (commonly with a cooling tower), some
water - and a lot of heat - will always be lost through evaporation.
For scope-2 water consumption, the focus is offsite. Less is to be said here, as similarly freshwater is used as
a cooling agent for large-scale power production, mostly at thermal and nuclear power plants. However, as
mentioned with AI becoming more and more desirable, the power that it uses will increase. In turn, the amount of
energy needing to be produced will rise as well, leading to more and more water usage.
6. (Ren, 2023)
For source, see (Ren, 2023) in References